Accurate classification of single nucleotide poly-morphisms (SNPs) is essential for genomic medicine. In this study, we fine-tune DNABERT, a transformer pre-trained on the human genome using k-mer embeddings for SNP pathogenicity prediction. Using a curated dataset of 50,000 Clin Var variants, our model achieves 94 % training accuracy, 93.4 % validation accuracy, an Fl-score of 0.938, and an AUC of 0.96. To improve interpretability, we integrate explainable AI (XAI) methods including Layer-Wise Relevance Propagation (LRP), Integrated Gradients, and DeepLIFT. These are evaluated using fidelity, completeness, robustness, and biological relevance metrics. LRP consistently delivers the most interpretable and biologically aligned attributions. Our findings show that combining fine-tuned DNABERT with effective XAI enables a transparent and scalable framework for SNP classification in clinical genomics.